Exploring the superiority of encoder-decoder architecture over traditional image processing techniques for binary segmentation of underwater images

• Main Authors: George, G., Anusuya, Kanakala, A., Lau, C. Y.
• Format: Conference Proceeding
• Language: English
• Subjects: Encoders-Decoders; Image processing; Image segmentation; Naval engineering; Performance evaluation; Recall; Sensitivity analysis; Underwater; Underwater exploration
• ISSN: 0094-243X; 1551-7616
• DOI: 10.1063/5.0229289

Applications like underwater exploration, marine life monitoring, and underwater item detection depend heavily on the ability to segment underwater images. However, underwater habitats’ varied sights and challenging qualities, such as reduced visibility, make proper segmentation challenging. In this study, the performance of an encoder- decoder architecture, especially the Unet model, is compared to that of conventional image processing methods for binary segmentation of underwater images. Performance evaluation is based on critical criteria, such as sensitivity (recall), F1 score, precision, and accuracy. The evaluation’s findings show that, among the methods examined, the Unet model has the highest sensitivity (95.16%). This demonstrates how well it works to locate bright areas in underwater photos. The Unet model also achieves a noteworthy F1 score of 86.71%, illustrating a favorable balance between precision and recall. While the Random Forest method’s F1 score is slightly lower at 78.80%, it still displays comparable precision (75.46%) and accuracy (79.45%) values. The Otsu Level set technique performs worse on all metrics. The encoder-decoder architecture of the Unet model, which efficiently absorbs and uses both local and global contextual information throughout the segmentation process, can be credited with the model’s better performance.